I planned to use ssd-mobilenet-V2, but read the meaning that it is not the best choice if you find small objects in the picture, which license plates are. I need real-time solution, which I could port to the mobile platform like Raspberry pi & Jetson Nano
Asked
Active
Viewed 33 times
1 Answers
0
Mobilenet is actually a good choice if you plan to port it to a mobile platform like Raspberry pi & Jetson Nano because it typically has less than 4 million trainable parameters. The issue of finding small objects in an image will be the same issue whether you are using Mobilenet or some other model. The issue comes down to the relationship between the number of pixels in the region of interest (ROI) (ie license plate) versus the total number of pixels in the image. For example if you have an image of size 200 X 200 you have 40,000 pixels. If the license plate say is 30 X 10 than you have 300 pixels in the region of interest. The ratio of 300/40000=.0075 indicates the model will have a difficult time of getting high accuracy from my experience. There are various techniques to deal with this one of which I believe is called sliding windows with which I am not very familar but I do know they present significantly higher computational cost. My suggestion is to build thee model and see how it does. I think the original image size will be important. The larger the better but again computational cost increases by the square of the image size.
Gerry P
- 7,662
- 3
- 10
- 20